Recommender System Using LLMs in E-commerce

This thesis focuses on the design, implementation, and evaluation of an advanced recommender system for an e-commerce environment. The proposed solution combines vector embeddings with large language models (LLMs) to address key limitations of traditional recommendation approaches—particularly low semantic quality of recommendations, the cold-start problem for new products, and insufficient use of textual information for content-rich products. The outcome is a production-deployed system for an e-commerce store based on a retrieval–ranking architecture. The retrieval layer operates over three embedding indexes: product, session, and transaction embeddings, while the ranking layer uses an LLM to produce the final ordering of candidates based on semantic relevance. The thesis describes the complete data architecture implemented on the Keboola platform, including data cleaning, embedding generation, pipeline orchestration, output validation, and quality monitoring. It includes thorough evaluation via A/B testing in a production environment, demonstrating statistically significant improvements in key metrics, an increase in click-through rate (CTR) and an improvement in conversion rate. Furthermore, the thesis discusses ethical and security aspects of using LLMs, focusing on privacy protection and hallucination prevention. The results confirm that combining embeddings with large language models is an effective and practically deployable approach to product recommendation in domains where semantic similarity is crucial.